optimizing StgPtr allocate (Capability *cap, W_ n)

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
2 messages Options
Reply | Threaded
Open this post in threaded view
|

optimizing StgPtr allocate (Capability *cap, W_ n)

Bulat Ziganshin-2
Hello Glasgow-haskell-users,

i'm looking a the https://github.com/ghc/ghc/blob/23bb90460d7c963ee617d250fa0a33c6ac7bbc53/rts/sm/Storage.c#L680

if i correctly understand, it's speed-critical routine?

i think that it may be improved in this way:

StgPtr allocate (Capability *cap, W_ n)
{
    bdescr *bd;
    StgPtr p;

    TICK_ALLOC_HEAP_NOCTR(WDS(n));
    CCS_ALLOC(cap->r.rCCCS,n);

/// here starts new improved code:

    bd = cap->r.rCurrentAlloc;
    if (bd == NULL || bd->free + n > bd->end) {
        if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
            ....
        }
        if (bd->free + n <= bd->start + BLOCK_SIZE_W)
            bd->end = min (bd->start + BLOCK_SIZE_W, bd->free + LARGE_OBJECT_THRESHOLD)
            goto usual_alloc;
        }
        ....
    }

/// and here it stops

usual_alloc:
    p = bd->free;
    bd->free += n;

    IF_DEBUG(sanity, ASSERT(*((StgWord8*)p) == 0xaa));
    return p;
}


i  think  it's  obvious - we consolidate two if's on the crirical path
into the single one plus avoid one ADD by keeping highly-useful bd->end pointer

further   improvements   may   include   removing  bd==NULL  check  by
initializing bd->free=bd->end=NULL   and   moving   entire   "if" body
into   separate   slow_allocate()  procedure  marked  "noinline"  with
allocate() probably marked as forceinline:

StgPtr allocate (Capability *cap, W_ n)
{
    bdescr *bd;
    StgPtr p;

    TICK_ALLOC_HEAP_NOCTR(WDS(n));
    CCS_ALLOC(cap->r.rCCCS,n);

    bd = cap->r.rCurrentAlloc;
    if (bd->free + n > bd->end)
        return slow_allocate(cap,n);

    p = bd->free;
    bd->free += n;

    IF_DEBUG(sanity, ASSERT(*((StgWord8*)p) == 0xaa));
    return p;
}

this  change  will  greatly simplify optimizer's work. according to my
experience   current  C++  compilers  are  weak  on  optimizing  large
functions with complex execution paths and such transformations really
improve the generated code


--
Best regards,
 Bulat                          mailto:[hidden email]

_______________________________________________
Glasgow-haskell-users mailing list
[hidden email]
http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
Reply | Threaded
Open this post in threaded view
|

Re: optimizing StgPtr allocate (Capability *cap, W_ n)

Edward Z. Yang
Hi Bulat,

This seems quite reasonable to me. Have you eyeballed the assembly
GCC produces to see that the hotpath is improved? If you can submit
a patch that would be great!

Cheers,
Edward

Excerpts from Bulat Ziganshin's message of 2014-10-14 10:08:59 -0700:

> Hello Glasgow-haskell-users,
>
> i'm looking a the https://github.com/ghc/ghc/blob/23bb90460d7c963ee617d250fa0a33c6ac7bbc53/rts/sm/Storage.c#L680
>
> if i correctly understand, it's speed-critical routine?
>
> i think that it may be improved in this way:
>
> StgPtr allocate (Capability *cap, W_ n)
> {
>     bdescr *bd;
>     StgPtr p;
>
>     TICK_ALLOC_HEAP_NOCTR(WDS(n));
>     CCS_ALLOC(cap->r.rCCCS,n);
>
> /// here starts new improved code:
>
>     bd = cap->r.rCurrentAlloc;
>     if (bd == NULL || bd->free + n > bd->end) {
>         if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
>             ....
>         }
>         if (bd->free + n <= bd->start + BLOCK_SIZE_W)
>             bd->end = min (bd->start + BLOCK_SIZE_W, bd->free + LARGE_OBJECT_THRESHOLD)
>             goto usual_alloc;
>         }
>         ....
>     }
>
> /// and here it stops
>
> usual_alloc:
>     p = bd->free;
>     bd->free += n;
>
>     IF_DEBUG(sanity, ASSERT(*((StgWord8*)p) == 0xaa));
>     return p;
> }
>
>
> i  think  it's  obvious - we consolidate two if's on the crirical path
> into the single one plus avoid one ADD by keeping highly-useful bd->end pointer
>
> further   improvements   may   include   removing  bd==NULL  check  by
> initializing bd->free=bd->end=NULL   and   moving   entire   "if" body
> into   separate   slow_allocate()  procedure  marked  "noinline"  with
> allocate() probably marked as forceinline:
>
> StgPtr allocate (Capability *cap, W_ n)
> {
>     bdescr *bd;
>     StgPtr p;
>
>     TICK_ALLOC_HEAP_NOCTR(WDS(n));
>     CCS_ALLOC(cap->r.rCCCS,n);
>
>     bd = cap->r.rCurrentAlloc;
>     if (bd->free + n > bd->end)
>         return slow_allocate(cap,n);
>
>     p = bd->free;
>     bd->free += n;
>
>     IF_DEBUG(sanity, ASSERT(*((StgWord8*)p) == 0xaa));
>     return p;
> }
>
> this  change  will  greatly simplify optimizer's work. according to my
> experience   current  C++  compilers  are  weak  on  optimizing  large
> functions with complex execution paths and such transformations really
> improve the generated code
>
_______________________________________________
Glasgow-haskell-users mailing list
[hidden email]
http://www.haskell.org/mailman/listinfo/glasgow-haskell-users